Kaggle Mars Pressure Competition's Public Leaderboard Uses Chronological, Not Random, Split

A competitor in a Kaggle challenge to reconstruct missing pressure readings from NASA's MEDA instrument on the Perseverance rover discovered that the public leaderboard uses the first half of test data in time order, not a random sample. The boundary falls after roughly 1.97 million rows, cutting inside sol 253 of the Mars mission timeline. This distinction matters significantly because the test window spans a period where atmospheric pressure drops by over 100 pascals, meaning public and private scores measure different seasonal regimes. The competitor had spent six weeks tuning constants, offsets, and blend weights against the public score, which effectively amounted to extrapolation rather than interpolation once the chronological split was confirmed. The finding reframes the entire model-tuning process, as optimizing against an early-season public score does not reliably predict performance on the late-season private score.